> ## Documentation Index
> Fetch the complete documentation index at: https://docs.wazoo.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# MemMachine

> Worlds vs MemMachine: a graph-and-SQL memory service against a curated fact ledger.

## At a glance

|                 | Worlds                        | [MemMachine](https://github.com/memmachine) |
| :-------------- | :---------------------------- | :------------------------------------------ |
| Data model      | RDF triples in a ledger       | Graph for episodes, SQL for profiles        |
| How facts enter | Import, patches, agent writes | Logged conversations and profile writes     |
| Retrieval       | Hybrid search + SPARQL        | Graph traversal + vectors                   |
| Provenance      | Fact-level ledger             | Graph-derived                               |
| Deployment      | Hosted beta or embedded       | Self-hosted or cloud                        |

## The philosophical difference

[MemMachine](https://github.com/memmachine) is an open-source memory service
that splits its model in two. Episodic memory, the relational context of
conversations, lives in a Neo4j graph, while profile data lives in SQL.
Retrieval walks the graph to reconstruct who said what to whom. It emphasizes
token efficiency and works across LLM providers, including AWS Bedrock.

Worlds stores relational structure too, but as RDF triples in an append-only
ledger. The difference is where structure comes from. MemMachine derives context
from conversation logs after the fact; Worlds accepts only facts you assert
deliberately. Every fact in Worlds carries a path to its source and a point on
the ledger's timeline.

The token-efficiency claim is worth weighing honestly. MemMachine reports using
fewer tokens than comparable services. Worlds answers the same question with a
different currency: exactness. A SPARQL query returns a deterministic result
rather than a retrieved context window.

## When Worlds fits

* Facts must be true against a source of record, not derived from conversation
  logs.
* You want to [query relationships](/worlds/query) deterministically instead of
  walking a reconstructed graph.
* You need to know when a fact changed and who asserted it.
* You want provider-agnostic LLM and embedding choices, including AWS Bedrock
  ([provider-agnostic](/comparisons/index)).

## When Worlds does not fit

* You want episodic memory reconstructed from conversation logs, with no fact
  curation. Worlds ingests unstructured text but does not extract graph facts
  automatically ([update](/worlds/update)).
* You want a memory service native to AWS operations, such as IAM-based access
  and Bedrock-managed storage.
* Context-window token cost is the binding constraint and you want the memory
  layer to minimize it.

## Coexistence

Run MemMachine for the episodic layer, what an agent talked about and how it
reacted, and Worlds for the durable facts the agent may act on. Both store
graphs, but the Worlds ledger stays the authoritative source because its facts
are curated and verifiable.